[Fix] GLM 4.7 + NVFP4 + MTP (#17166)

python/sglang/srt/model_loader/loader.py

@@ -96,6 +96,7 @@ from sglang.srt.model_loader.weight_utils import (
     get_quant_config,
     gguf_quant_weights_iterator,
     initialize_dummy_weights,
+    maybe_add_mtp_safetensors,
     multi_thread_pt_weights_iterator,
     multi_thread_safetensors_weights_iterator,
     np_cache_weights_iterator,
@@ -321,6 +322,9 @@ class DefaultModelLoader(BaseModelLoader):
         fall_back_to_pt: bool = True
         """Whether .pt weights can be used."""
 
+        model_config: Optional["ModelConfig"] = None
+        """The model configuration (for checking architecture, etc)."""
+
         @classmethod
         def init_new(cls, model_config: ModelConfig, model):
             return cls(
@@ -328,6 +332,7 @@ class DefaultModelLoader(BaseModelLoader):
                 model_config.revision,
                 prefix="",
                 fall_back_to_pt=getattr(model, "fall_back_to_pt_during_load", True),
+                model_config=model_config,
             )
 
     def __init__(self, load_config: LoadConfig):
@@ -471,6 +476,15 @@ class DefaultModelLoader(BaseModelLoader):
         hf_folder, hf_weights_files, use_safetensors = self._prepare_weights(
             source.model_or_path, source.revision, source.fall_back_to_pt
         )
+
+        if use_safetensors and source.model_config is not None:
+            hf_weights_files = maybe_add_mtp_safetensors(
+                hf_weights_files,
+                hf_folder,
+                "model.safetensors.index.json",
+                source.model_config.hf_config,
+            )
+
         if self.load_config.load_format == LoadFormat.NPCACHE:
             # Currently np_cache only support *.bin checkpoints
             assert use_safetensors is False

python/sglang/srt/model_loader/weight_utils.py

@@ -594,6 +594,44 @@ def filter_duplicate_safetensors_files(
     return hf_weights_files
 
 
+def maybe_add_mtp_safetensors(
+    hf_weights_files: List[str], hf_folder: str, index_file: str, hf_config
+) -> List[str]:
+    """
+    Auto-detect and add mtp.safetensors for GLM4Moe MTP/NextN models if:
+    1. mtp.safetensors exists in the model directory
+    2. mtp.safetensors is NOT in the index (checkpoint packaging bug)
+    3. Model architecture is Glm4MoeForCausalLM with num_nextn_predict_layers > 0
+
+    This works around incorrectly packaged FP4 checkpoints like
+    baseten-admin/glm-4.7-fp4 where mtp.safetensors exists but
+    isn't referenced in model.safetensors.index.json.
+    """
+    # Only apply for GLM4Moe architecture with nextn layers
+    arch = getattr(hf_config, "architectures", [None])[0]
+    num_nextn_layers = getattr(hf_config, "num_nextn_predict_layers", 0)
+    if not (
+        arch in ["Glm4MoeForCausalLM", "Glm4MoeForCausalLMNextN"]
+        and num_nextn_layers > 0
+    ):
+        return hf_weights_files
+
+    # Check if mtp.safetensors exists and is not already in the file list
+    mtp_path = os.path.join(hf_folder, "mtp.safetensors")
+    if not os.path.isfile(mtp_path) or mtp_path in hf_weights_files:
+        return hf_weights_files
+
+    # mtp.safetensors exists but not in index - this is a bug
+    logger.warning(
+        f"Found mtp.safetensors but it's not referenced in {index_file}. "
+        f"This is a checkpoint packaging bug. Auto-adding it for loading. "
+        f"Please report this to the checkpoint provider."
+    )
+
+    # Add it to the files list
+    return hf_weights_files + [mtp_path]
+
+
 def filter_files_not_needed_for_inference(hf_weights_files: List[str]) -> List[str]:
     """
     Exclude files that are not needed for inference.